A Defensive Agent Approach to Improve Byzantine Robustness of Distributed Learning Systems

Abstract

The invention belongs to the technical field of distributed artificial intelligence, and specifically relates to a defense agent method for improving the Byzantine robustness of a distributed learning system. The present invention uses the adaptive credibility evaluation module based on the neural network structure to dynamically evaluate the credibility of each submission gradient, update the global classifier parameters maintained on the current master node, generate reward signals, and use the reward signals in reinforcement learning. Adjust the parameters of the adaptive credibility evaluation module under the framework; dynamically adjust the feasibility evaluation value of each working node during the training process, and alleviate the impact of the tampered gradient submitted by malicious working nodes on the system training process, so as to improve the distributed Byzantine robustness of learning systems. The present invention can be widely applied to various distributed deep learning systems, and improves the Byzantine robustness of the system. The security of the distributed training process of the artificial intelligence system has been significantly improved, especially when the proportion of malicious working nodes is greater than or equal to 50%.

Description

technical field [0001] The invention belongs to the technical field of distributed artificial intelligence, and in particular relates to a defense agent method for improving the Byzantine robustness of a distributed learning system. Background technique [0002] With the continuous development of deep learning technology, massive training data is being put into the training process of classifiers based on machine learning, and building classifiers on distributed platforms has gradually become a mainstream trend in the industry. However, due to factors such as transmission errors and deliberate tampering, some working nodes in the distributed learning system may submit tampered gradients to the master node to induce the master node's gradient merging algorithm to be abnormal, resulting in wrong parameter update decisions and disrupting the normal operation of the classifier. The training process, which is often referred to as a Byzantine attack on distributed learning systems...

AI Technical Summary

Problems solved by technology

However, due to factors such as transmission errors and deliberate tampering, some working nodes in the distributed learning system may submit tampered gradients to the master node to induce the master node's gradient merging algorithm to be abnormal, resulting in wrong parameter update decisions and disrupting the normal operation of the classifier. The training process, which is often referred to as a Byzantine attack on distributed learning systems

Most of the existing Byzantine robustness improvement methods rely on the "majority voting mechanism" between gradients. Once most of the working nodes in the system experience Byzantine anomalies or are maliciously controlled, these defense mechanisms will fail and the distributed learning system will not be able to obtain orders. Satisfactory Byzantine Robustness

Images